reexport_vits_with_controls.py

import argparse
import shutil
import subprocess
from pathlib import Path

import onnx
import torch
from transformers import AutoModel, AutoTokenizer


class VitsExportWrapper(torch.nn.Module):
    def __init__(self, model: torch.nn.Module):
        super().__init__()
        self.model = model.eval()

    def forward(
        self,
        input_ids: torch.Tensor,
        attention_mask: torch.Tensor,
        speaker_id: torch.Tensor,
        emotion_id: torch.Tensor,
    ) -> torch.Tensor:
        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            speaker_id=speaker_id.to(torch.long),
            emotion_id=emotion_id.to(torch.long),
            return_dict=True,
        )
        return outputs.waveform


def inspect_onnx(onnx_path: Path) -> None:
    model = onnx.load(str(onnx_path))
    print("onnx inputs:")
    for value in model.graph.input:
        tensor_type = value.type.tensor_type
        dims = []
        for dim in tensor_type.shape.dim:
            if dim.dim_value:
                dims.append(dim.dim_value)
            elif dim.dim_param:
                dims.append(dim.dim_param)
            else:
                dims.append("?")
        print(f"  {value.name}: shape={dims}, elem_type={tensor_type.elem_type}")


def inspect_mnn(mnn_path: Path) -> None:
    import MNN.expr as expr

    graph_vars = expr.load_as_dict(str(mnn_path))
    for name in ("input_ids", "attention_mask", "speaker_id", "emotion_id"):
        if name not in graph_vars:
            print(f"mnn input missing: {name}")
            continue
        var = graph_vars[name]
        print(f"mnn input {name}: shape={var.shape}, dtype={var.dtype}, format={var.data_format}")


def export_onnx(args: argparse.Namespace) -> None:
    model_dir = Path(args.model_dir)
    tokenizer = AutoTokenizer.from_pretrained(model_dir, trust_remote_code=True, local_files_only=True)
    model = AutoModel.from_pretrained(model_dir, trust_remote_code=True, local_files_only=True)
    wrapper = VitsExportWrapper(model)

    tokenized = tokenizer(text=args.text, return_tensors="pt")
    input_ids = tokenized["input_ids"].to(torch.long)
    attention_mask = tokenized.get("attention_mask", torch.ones_like(input_ids)).to(torch.long)
    speaker_id = torch.tensor([args.speaker_id], dtype=torch.long)
    emotion_id = torch.tensor([args.style_id], dtype=torch.long)

    torch.onnx.export(
        wrapper,
        (input_ids, attention_mask, speaker_id, emotion_id),
        str(args.onnx_output),
        input_names=["input_ids", "attention_mask", "speaker_id", "emotion_id"],
        output_names=["waveform"],
        dynamic_axes={
            "input_ids": {1: "text_length"},
            "attention_mask": {1: "text_length"},
            "waveform": {1: "audio_length"},
        },
        opset_version=args.opset,
        do_constant_folding=True,
        dynamo=False,
    )
    print(f"wrote {args.onnx_output}")
    inspect_onnx(Path(args.onnx_output))


def convert_to_mnn(args: argparse.Namespace) -> None:
    if not args.mnn_output:
        return

    converter = shutil.which("MNNConvert")
    if converter is None:
        raise FileNotFoundError("MNNConvert not found in PATH")

    command = [
        converter,
        "-f",
        "ONNX",
        "--modelFile",
        str(args.onnx_output),
        "--MNNModel",
        str(args.mnn_output),
        "--bizCode",
        "MNN",
    ]
    subprocess.run(command, check=True)
    print(f"wrote {args.mnn_output}")
    inspect_mnn(Path(args.mnn_output))


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description="Re-export local VITS weights with speaker/style control inputs.")
    parser.add_argument("--model-dir", default=".", help="Directory containing config, tokenizer, custom code, and weights.")
    parser.add_argument("--text", default="வணக்கம்", help="Sample text used to trace the export graph.")
    parser.add_argument("--speaker-id", type=int, default=18, help="Sample speaker ID used during tracing.")
    parser.add_argument("--style-id", type=int, default=0, help="Sample style or emotion ID used during tracing.")
    parser.add_argument("--onnx-output", default="vits_tamil_with_controls.onnx", help="Path for exported ONNX.")
    parser.add_argument("--mnn-output", default="vits_tamil_with_controls.mnn", help="Optional output path for converted MNN.")
    parser.add_argument("--opset", type=int, default=17, help="ONNX opset version.")
    return parser.parse_args()


def main() -> None:
    args = parse_args()
    export_onnx(args)
    convert_to_mnn(args)


if __name__ == "__main__":
    main()